Page 141 - Proceeding of Atrans Young Researcher's Forum 2019_Neat

“Transportation for A Better Life:
Smart Mobility for Now and Then”

23 August 2019, Bangkok, Thailand

Figure 5. Error between target and values of RMSE, MAE, and accuracy evaluated for
predicted values for the testing dataset of: the statistical analysis of SVM and EDT Bagged
(a) SVM algorithm and (b) EDT Bagged algorithms.
Figure 6 shows the statistical results over 1000
From the obtained results, it can be concluded Monte Carlo simulations of RMSE, MAE and
that both SVM and EDT Bagged algorithms are accuracy for EDT Bagged and SVM algorithms. It is
potential candidates for predicting the travel observed that both algorithms performed well the
decisions of transport users. However, EDT Bagged prediction task of the problem. The statistical results
yielded a slightly better result than SVM. were quite stable, as the mean values of RMSE=0.88
for EDT Bagged and RMSE=0.89 for SVM, the
3.2. Robustness of AI Algorithms mean values of MAE=0.35 for EDT Bagged and

In order to investigate the robustness of the MAE=0.37 for SVM, whereas that of accuracy=0.80
proposed AI models, 1000 Monte Carlo simulations, for EDT Bagged and accuracy=0.78 for SVM. The
which is presented by Rubinstein [63], have been corresponding standard deviation were 0.0714,
performed by each proposed AI method. In classical 0.0434 and 0.0219 for RMSE, MAE and accuracy of
transport survey, uncertainty sources might come EDT Bagged, and 0.0716, 0.0442, 0.0223 for
from the false responses of travel users leading to RMSE, MAE and accuracy of SVM. A small
incorrect data values and responses. As regard to the variation of the error criteria clearly showed the
AI modeling part, the selection of samples to performance and efficiency of the prediction tools in
construct the 70% training and 30% testing dataset solving such high dimensional input space
might affect the predicted output results. The reason classification problem (15 inputs). The effect of the
that Monte Carlo approach was chosen as it is a very choice of sample to construct the 70% training and
efficient method to propagate uncertainties of the 30% testing dataset can be seen in the work of Dao
input to the output space. Parallel computing was et al. [64], where the ANN method is very unstable
performed to obtain a number of 1000 corresponding compared to ANFIS.

116